Вы находитесь на странице: 1из 9


6, JUNE 2002 565

Movement Quantification in Epileptic Seizures:

A New Approach to Video-EEG Analysis
Zhanjian Li*, António Martins da Silva, and João Paulo Silva Cunha, Member, IEEE

Abstract—It is common that epileptic seizures induce uncoor- of the seizure to be made. To our knowledge, no effort has been
dinated movement in a patient’s body. This movement is a rele- made to develop quantification algorithms that extract informa-
vant clinical factor in seizure identification. Nevertheless, quan- tion from video recorded seizures. The main reasons for this are,
tification of this information has not been an object of much at-
on the one hand, the traditional video monitoring units rely on
tention from the scientific community. In this paper, we present
our effort in developing a new approach to the quantification of analogs video recording devices, which are an obstacle to the
movement patterns in patients during epileptic seizures. We attach use of image analysis approaches and, on the other hand, cap-
markers at landmark points of a patient’s body and use a camera turing, storing, and processing video data demands enormous
and a commercial video-electroencephalogram (EEG) system to throughput from computers. Recently, the emergence of full dig-
synchronously register EEG and video during seizures. Then, we ital video-EEG technology based on the support of computer
apply image-processing techniques to analyze the video frames and multimedia has opened up such possibilities.
extract the trajectories of those points that represent the course of
the quantified movement of different body parts. This information We developed a new approach to quantify patient movement
may help clinicians in seizure classification. We describe the frame- during epileptic seizures based on a commercial video-EEG
work of our system and a method of analyzing video in order to system. Our motivation was propted by the following two facts.
achieve the proposed goal. Our experimental results show that our • Clinicians consider uncoordinated seizure motion pattern
method can reflect quantified motion patterns of epileptic seizures,
(or motor semiology) of the patient’s body as valuable ev-
which cannot be accessed by means of traditional visual inspection
of video recordings. We were able, for the first time, to quantify idence in the identification of the presence of a seizure
the movement of different parts of a convulsive human body in the and of its source [1], [3]. Traditional visual inspection of
course of an epileptic seizure. This result represents an enhanced long-term video-EEG recordings cannot evaluate quanti-
value to clinicians in studying seizures for reaching a diagnosis. fied motion information.
Index Terms—Epilepsy, movement quantification, video-EEG. • Computer technology is in constant development. Dig-
ital video systems and video-EEG equipment are being
applied progressively more in the clinical diagnosis of
These facts have enabled us to promote a line of research that
I T IS common that epileptic seizures induce uncoordinated
movement in a patient’s body. This movement is a relevant
clinical factor in seizure identification. Nevertheless, quantifi-
aims to study movement patterns of epileptic seizures, by means
of computer technology. From the clinical point of view, this
cation of this information has not been an object of much at- research addresses the following questions.
tention from the scientific community. Traditionally, long-term • Can quantified movement analysis extract information
video-electroencephalogram (EEG) monitoring is employed to that is imperceptible to the visual inspection of an epilep-
study these cases [1], [2]. Although movement is very important tologist?
in the diagnosis of epilepsy, for many years, researchers have • Can movement quantification information be employed
concentrated on developing quantification algorithms to extract to provide additional evidence in the classification of
EEG features which are invisible to traditional visual inspection. epileptic seizures?
Similar approaches have not been carried out on video informa- • What kind of quantified movement information is helpful
tion, where clinicians only perform visual inspection. This vi- in the diagnosis of epilepsy?
sual analysis allows only a rough estimate of the motion pattern • Do these techniques contribute toward a more reliable di-
Manuscript received July 10, 2001; revised December 12, 2001. This work In this paper, we propose a new experimental approach to start
was supported by the Portuguese Science and Technology Foundation (F.C.T.) addressing these questions.
under PRAXIS XXI and POSI programs, grants BD/13603/97 and Sapiens
33802/99. Asterisk indicates corresponding author.
*Z. Li is with the Department of Electronics and Telecommunication/IEETA, II. RELATED WORK
University Campus of Santiago, University of Aveiro, 3810-193 Aveiro, Por-
tugal (e-mail: li@ieeta.pt). Extracting human motion has been an important research
A. Martins da Silva is with the Neurophysiology Department, S. António
General Hospital, 4099-001 Porto, Portugal. He is also with ICBAS, Univer-
topic in computer vision for many years. However, analyzing
sity of Porto, 4099-001 Porto, Portugal. the motion of the human body is a challenging problem because
J. P. Silva Cunha is with the Department of Electronics and Telecommu- the human body is a highly flexible structure that can twist
nication/IEETA, University of Aveiro, 3810-193 Aveiro, Portugal. He is also and bend in many degrees of freedom. In addition, the variety
with the Neurophysiology Department, S. António General Hospital, 4099-001
Porto, Portugal (e-mail: jcunha@det.ua.pt). of luminance, noise induced by devices and self-occlusions of
Publisher Item Identifier S 0018-9294(02)04846-2. the human body complicate the movement analysis process.
0018-9294/02$17.00 © 2002 IEEE

To address this complex problem, some investigators used every center. Usually, a patient lies on a bed and is monitored
marker-based methods to evaluate human body motion for by a video camera and an EEG machine. Our experimental
clinical diagnosis. For instance, Jobbágy et al. [4] used a setup is illustrated in Fig. 1.
marker-based method to analyze movement patterns of a pa- A monochrome CCD camera with several infrared light emit-
tient hand for aiding in early detection of Parkinson’s disease. ting diodes parallel to the lens are mounted right above the bed
These authors attached markers to the middle phalanges of a at an approximate distance of 1.5 m. A patient lies on the bed.
patient’s hand and extracted marker trajectories for evaluation. The image size is adjusted to bring the whole body of the sub-
In their experiments, the subject was requested to put his hands ject into focus. A set of infrared reflective markers is attached to
on the desk and pinch, twiddle, and tap. A camera captured landmark points of the body. This is done through black elastic
this procedure and the vertical displacements of markers were braces where these markers are attached in order to enhance the
extracted and evaluated. Miyazaki et al. [5] developed a system contrast in the image. The light sources in the ceiling are in-
to quantify the eyelid movement for clinical purposes. They candescence lamps which are kept at relatively low intensity in
attached markers to the lower margin of the upper eyelid of order that highlighted blobs are formed in the video image only
a subject. The subject was requested to blink and the marker by infrared reflective markers.
movements were recorded and evaluated. Motion quantification We assume that the motion brought on by the seizure cannot
can also be performed without markers. Nakajima et al. [6] cause the patient to fall out of view, such as falling off the bed.
used an image optical flow analysis approach to quantitatively Hence, the camera is kept stationary without tilting or rotating
evaluate body motion and visualize the apparent velocity field during the video capture. A commercial video-EEG system [11]
of the entire body motion, including both chest respiratory is used to synchronously acquire video sequences and EEG data
movement and posture change. into a PC. The video data is saved in a video file that can then be
Although these approaches are suitable for their target appli- analyzed by means of a software tool we have developed, called
cations, they cannot be used to extract quantified information quantification of movement in epileptic seizures (QMovES).
for more complex human motion. For instance, in the work car-
ried out by Jobbágy et al. [4], the subject’s hands are required B. Marker Definition
to be placed steadily on the desk so that the relating positions of
Due to the originality of our approach and to the consequent
the markers (fingers) do not change, and the markers are not oc-
lack of established standards, we defined a marker positioning
cluded. This constraint is very important because the highlight
system with 22 landmark points (MPS22) on the human body
markers are easily identified according to their horizontal po-
as illustrated in Fig. 2.
sitions to enable the establishment of motion correspondence.
The markers should be attached to the landmark points be-
However, this approach will fail in complex movements such as
fore the video acquisition. They are covered with infrared re-
rotation, crossing, etc. The approaches that are based in optical
flow, of which the work of Nakajima et al. [6] is an example, flective material so that they reflect this light invisible to human
only work in the case of simple motion due to their high sensi- vision. They are very small and lightweight not blocking the
tivity to noise. patient movement. An infrared-sensitive CCD camera shows
Some researchers used model-based approaches for tracking the markers as high intensity clusters of pixels (blobs) in video
or recognizing human motion such as walking [7], [8] or other frames that are detectable through our method.
simple motions [9], [10]. However, these approaches can only
C. Extraction of Quantified Information of Motion
work for a few simple human motions. It is difficult to use
them in our target application because the movement induced QMovES imports the video file as input of its quantified
by epileptic seizures is complex and uncoordinated and is, thus, movement extracting method. This method is divided in two
very difficult to model and quantify. From our literature review, stages: initialization and tracking, as illustrated in Fig. 3.
marker-based methods seemed to be the most suitable for our 1) Initialization: This stage includes calculating a
targeted application. threshold, segmenting, and assigning labels to blobs as
In order to attain our primary goal, we attach some markers well as calculating initial parameters used in the tracking stage.
to the landmark points of a patient’s body so that the motion Due to the framework we are using, the intensity of pixels
of the markers reflects the movement of underlying body parts. of highlighted blobs is higher than pixels of other objects in
A method is proposed to obtain quantified information of com- the scene. To segment them, we calculate a grey level threshold
plex human motion. We focus our attention on knowing which from a frame based on an approach purposed by Johannsen [12]
parts of the human body move and what are their dynamics. This using an entropy measure. This threshold is used to segment the
strategy simplifies the procedure of extracting motion patterns blobs in video frames in the tracking stage, assuming that the
by avoiding the need to address topics such as gesture recogni- indoor luminance does not change too much during video-EEG
tion or geometry restructure. acquisition. After segmentation, we calculate the geometric cen-
troid of the blobs. In this step, we assume that clustered pixels
III. METHODS in a small search window belong to one blob. This constraint
is intuitively similar to human perception. A scan is performed
A. The Framework to find all the blobs in the image. Given our application do-
In order to ensure clinical acceptance, we designed our main, we cannot assume what the patient’s initial posture was
approach according to the routine video-EEG procedure used and even cannot assume that all defined markers are exposed
in clinical neurophysiology environments, which is similar in to view. Thus, it is very difficult to establish a complete marker

Fig. 1. The setup of QMovES system. A CCD camera acquires video sequences into a PC where a digital video-EEG system and the QMovES software tool are

In our method we model the movement of a marker as a two-

dimensional (2-D) ballistic motion [13]. A marker position in
two successive frames can be approximately expressed by
where and denote the centroid coordinate of the marker,
and denote instantaneous velocity along the axis direc-
tions, and the corresponding instantaneous accelerations,
is frame number, and denotes a small temporal sampling
interval that is equal to 1/(frame rate). Assuming we have the
three adjacent frames , , and , then we can estimate
and and and as follows:


Fig. 2. The 22 landmark positions we defined for the full-body marker (5)
positioning system (MPS22).

positioning system necessary for automatic labeling, because no
prior knowledge about the patient’s posture is available for ini- Given this model, we use the first three frames of seizure
tialization and training purposes. Given these facts, we perform recordings to calculate the initial parameters according to
the initial labeling procedure interactively by means of a mouse (3)–(6) which are necessary to initialize a Kalman filter [14]
click sequence. for prediction purposes in the tracking phase.

(a) (b)
Fig. 4. Illustrations of the “correspondence problem” where the point’s
motion from one frame to next are represented. The opened circles denote
the points in the first frame and the filled circles denote the points in the
next frame. The squares are predicted positions for the points. (a) Overlap of
two frames to show the positions change. (b) Denotes actual correspondence
solution computed (arrows).

where is the number of markers, the number of blobs

found in the current frame, and is the pre-
dicted position of th marker; is
the position of th blob found, denotes the magnitude of
vector , and denotes the th frame. The blobs are identified
as correspondent to respective markers by minimizing the sum
Fig. 3. Schematic representation of the procedure for movement quantification of the normalized distance between the predictive and the ef-
used in QMovES.
fective positions of the blobs. This normalized distance already
takes into consideration speed, direction and acceleration be-
2) Tracking: The tracking stage of our method uses the cause marker positions are predicted by (1) and (2). The de-
information obtained from the initialization stage, and iterates nominator denotes the sums of absolute quantities for all pos-
it for every remaining frame in the sequence to compute the sible matches. Hence, the proposed measure gives the relative
successive positions of the markers. In this procedure, a video measure of distance from predicted position to actual position.
frame is first extracted and segmented as described previously. Large displacements will lead to large values.
Then the clusters of highlight pixels (blobs) are extracted using Although this proposed solution proved to work correctly, the
the algorithm proposed by Jobbágy [15]. Following this, marker correspondence problem is combinatorial explosive. If there are
positions are predicted and used in the motion correspondence points in a sequence of frames, resulting in trajectories,
estimation section of the method. This is a key part of our the number of possible trajectory sets is which may
method. To provide the reader with a better understanding of lead to a highly complex and computational intensive solution.
this problem and of the solution we have designed for it, we To cope with this complexity we implemented in QMovES a
present an illustration of a situation where this problem occurs modified version of a noniterative Greedy algorithm originally
in Fig. 4. We assume the points (labeled with A, B, C, D, and E) proposed by Rangarjan et al. [16] where (7) is used as cost func-
move at different velocity and acceleration. The opened circles tion. This algorithm deals with video frame by frame in a loop,
denote the points in one frame, which are already labeled, and in which the outputs are the positions of markers in the cur-
the filled circles denote the points in the next frame. We overlap rent frame. These outputs are imported into the Kalman filters
two frames in one figure to show how the position changes. The to iteratively predict parameters for the next loop. Through this
problem at this point is how we label the filled circle points in procedure, the correspondence is solved as shown in Fig. 4(b).
Fig. 4(a) and is technically named “correspondence problem” In our application domain the set of positions of the markers
[16]. (trajectories) reflect the quantified motion information of the
As referred previously, we model the movement by (1) and (2) underlying body parts. The movement patterns can then be ex-
and use a Kalman filter to predict marker positions in the next tracted from this information.
video frame. These predicted positions, presented in Fig. 4 as
squares, are then used in motion correspondence estimation by IV. RESULTS
a distance criteria optimization algorithm based on a normalized
distance that we define as Fig. 5(a) is an original video frame taken under our frame-
work in which the highlight blobs on the frame denote the
markers attached to the patient’s body. Fig. 5(b) shows the
(7) segmentation result where the blobs are extracted from the
background. Fig. 5(c) shows the centroid positions of blobs
(cross markers) found by the algorithm described in the

(a) (b)

(c) (d)
Fig. 5. (a) An original video frame with highlight blobs, (b) the segmentation result, (c) detection and correspondence of the blobs, (d) the output of tracking
stage with predicted positions ( ) and actual positions ( ).

previous section. Fig. 5(d) shows an output of a tracking pro- This type of visualization technique is commonly used in neu-
cedure. The squares denote the predicted positions of markers rophysiology to reconstruct the underlying potentials of EEG or
generated by the Kalman filter. The crosses denote the actual their specific frequency band’s power distribution in the scalp.
positions of markers located by the motion correspondence Once more, it was our intention to ease the clinicians’ adapta-
estimation part of our method. tion to the reading of this new type of information by using a
The output trajectories of those markers from this video se- familiar approach to present it.
quence are showed in Fig. 6. Fig. 6(a) shows the motion of the In Fig. 7 we show several frames of an animated color-coded
markers (pixel unit) in axis direction and Fig. 6(b) shows sim- graphic of quantified movement processed by a median filter.
ilar information in the axis direction. In Fig. 6(c), the speed The highlighted areas represent the intensity of movement of
of the markers is depicted. If we know the mass of the under- those body parts. The scale in the left bar denotes the motion
lying body parts, this data can be converted to show the effective speed in pixels/s. From Fig. 7, we can learn the motion dynamic
quantity of movement involved in each section of the convulsive pattern during the seizure and characterize its evolution. For in-
body. stance, we can quantify the hand’s motion at time 00:06:280
We decided to display the movement-quantified information and the leg’s intense motion at times 00:07:120, 00:07:360, and
by mimicking EEG tracings in order to ease the adaptation of 00:10:000. We can also analyze the intensity of the head move-
the clinician to the reading this information and to facilitate the ment observed at times 00:10:000 and 00:20:760. This infor-
correlation between EEG and movement events in an easy way. mation is made available as a video sequence to dynamically
The marker trajectories are then the quantified motion informa- visualize the motion pattern exhibited by the patient during a
tion, which denotes the motion pattern of different body parts. seizure.
For instance, from Fig. 6(c), we can learn that the legs (K1, K2,
J1, J2, I1, and I2) move intensively from second 6 to second 21.
In Fig. 6(b), from the tracings of A1, A2, B1, and B2, we can
measure the behavior of the head that shows six rotations from Given the originality of the approach we have presented, we
second 5 to second 21. One important achievement is that this are not able to compare it with other studies in the literature. We
information is given to the user in a quantified way. We are able can identify strong advantages in its application but also have
to register movement events and quantify their speed and spatial to face several problems that should occupy our efforts in the
motion patterns. This type of quantification may make a lot of course of this line of research in the near future.
difference in clinical application and is apparently achieved for The digital video-EEG technique is a high input–output
the first time. throughput application which is characterized by large data
To enhance the pattern recognition capability, we also visu- sets that usually are only bearable by computer systems when
alize the motion pattern by an animated color-coded graphic. applying high compression to the original data sets. However,


Fig. 6. (a) Tracings of the markers motion in coordinate in function of time with the corresponding body positions. (b) Tracings of the markers motion in Y
coordinate in function of time with the corresponding body positions.

a high compression rate leads to a loss of image information cause difficulties in image segmentation. For robust image
and degradation in the quality of the image. In practice, to save segmentation we used dark colored background objects and
disk space, the compression rate should be kept high while low-medium luminance. We also recommended that patients
video quality goes to a relatively low (but acceptable) level. wear dark colored clothes if possible.
In our study, the video data is saved in M-JPEG format with a A difficulty in the tracking procedure of our approach arises
compression rate of 16 : 1. Our approach proved to be highly when some markers may become hidden from view due to oc-
robust to compression noise and can work correctly with rela- clusion by other body parts during the movement. This case is
tively low-quality images. In our system setup, the intensity of more serious if the patient lies sideways or in a prone position on
the markers is almost uniform for different indoor luminance. the bed. In these situations there is no prediction model available
Under the condition of low luminance, the intensity of a diffuse and the marker detection algorithm cannot work properly. Nev-
background is always lower than that of the markers even if ertheless, in cases of temporary occlusion, we can visually esti-
the patient is wearing white clothes. In this case, the algorithm mate positions and assign labels for them manually. This work
can work properly. However, when luminance is increased, the can be done simply by mouse click. The benefit of this proce-
luminosity of the background will also increase, which may dure is that the occluded points in the frame will have approxi-

Fig. 6. (Continued.) (c) Tracings of the markers’ speed in function of time with the corresponding body positions.

Fig. 7. Frames from an animated color-coded video showing the dynamics of the quantity of movement in the course of a seizure. This procedure shows very
impressive results in the perception of movement patterns.

mately correct positions that still contribute to the trajectories in markers attached to the face of the subject will never change
the final output. However, this method is not valid for the cases their relative positions. As those approaches have proven to
of long-term occlusion. be quite effective, we have inserted some heuristic rules in
The markers may be lost during the tracking phase if more our method. If the heuristic automated decision does not solve
than two markers move too close, although this situation these problems, the process may be paused (automatically
is relatively rare when compared with occlusion. QMovES or manually) and the user may identify the blobs by simple
may also label markers wrongly if a large displacement or an mouse click. Automatic labeling can then continue after this
abrupt change in velocity is registered. These cases violate the interactive interruption.
correspondence assumptions so that the optimal match does not One limitation of QMovES is the lack of deepness infor-
correspond to the correct one. Some heuristic approaches may mation. However, our experiments showed that our quantified
also be used to respond to these problems. For instance, the four movement information extracted from 2-D images is still able

to reflect the motion pattern of the human body. One other in- ACKNOWLEDGMENT
convenience is that we must attach markers to the patient, which The authors acknowledge the two awards this work has al-
is an additional task to standard clinical video-EEG acquisition ready received from the European Epilepsy Academy and the
procedure. To simplify the procedure of attaching markers, we Portuguese League Against Epilepsy. They would also like to
fixed the markers to elastic braces. These elastic braces can be thank Dr. T. Roberto for the revision of the manuscript.
attached to body parts in a few minutes. In mild seizures the
markers usually do not change their relative positions according
to the MPS22 model of Fig. 2. Nevertheless, in some more vi-
[1] L. F. Quesney and P. Gloor, “Localization of epileptic foci,” in
olent ones, this may happen. We are already considering other Long-Term Monitoring in Epilepsy (EEG Suppl. no. 37), J. Gotman, J.
solutions to avoid this problem. R. Ives, and P. Gloor, Eds., 1985.
[2] M. C. Salinsky, “A practical analysis of computer based seizure de-
tection during continuous video-EEG monitoring,” Electroencephalogr.
Clin. Neurophysiol., vol. 103, pp. 445–449, 1997.
[3] H. Lüders, J. Acharya, C. Baumgartner, S. Benbadis, A. Bleasel,
R. Burgess, D. S. Dinner, A. Ebner, N. Foldvary, E. Geller, H.
VI. CONCLUSION AND FUTURE WORK Hamer, H. Holthausen, P. Kotagal, H. Morris, H. J. Meencke, S.
Noachtar, F. Rosenow, A. Sakamoto, B. J. Steinhoff, I. Tuxhorn, and
E. Wyllie, “Semiological seizure classification,” Epilepsia, vol. 39, pp.
This work has investigated an original approach to extract 1006–1013, 1998.
quantified motion information in the course of epileptic [4] Á. Jobbágy, E. H. Furnée, P. Harcos, M. Tárczy, I. Krekule, and L.
Komjáthi, “Analysis of movement patterns aids the early detection of
seizures. We attached markers to patients’ bodies and per- Parkinson’s disease,” in Proc. 19th IEEE/EMBS, Chicago, IL, Oct.
formed the video-EEG procedure in a standard clinical 30–Nov. 2 1997, pp. 1760–1763.
[5] S. Miyazaki, A. Ishida, and A. Komatsuzaki, “A clinically oriented
neurophysiology environment. A method to extract the trajec- video-based system for quantification of eyelid movement,” IEEE
tories of markers is also presented. Our approach has proved to Trans. Biomed. Eng., vol. 47, pp. 1088–1096, Aug. 2000.
be able to extract quantified motion trajectories of human body [6] K. Nakajima, A. Osa, T. Maekawa, and H. Miike, “Evaluation of body
motion by optical flow analysis,” Jpn. J. Appl. Phys., pt. 1, vol. 36, no.
parts that reflect the motion pattern of patients during epileptic 5A, pp. 2929–2937, 1997.
seizures. These trajectories can be reviewed as multichannel [7] D. Hogg, “Model-based vision: A program to see a walking person,”
tracings which are similar to EEG. Through further processing, Image Vis. Computing, vol. 1, no. 1, pp. 5–20, 1983.
[8] H. J. Lee and Z. Chen, “Knowledge-guided visual perception of 3-D
we can visualize the motion pattern in the form of an animated human gait from a single image sequence,” IEEE Trans. Syst., Man, Cy-
color-coded graphic. These ways of visualizing seizure motion bern., vol. 22, pp. 336–342, Mar.-Apr. 1992.
are also original and proved to be very effective in the clinical [9] K. Akita, “Image sequence analysis of real word human motion,” Pattern
Recogn., vol. 17, no. 1, pp. 73–83, 1984.
trials we have performed. Our experimental results show [10] C. R. Wren, A. Azarbayejani, T. Darrell, and A. P. Pentland, “Pfinder:
that the quantified motion information intuitively reflects the Real-time tracking of the human body,” IEEE Trans. Pattern Anal. Ma-
movement patterns of epileptic seizures. These quantified chine Intell., vol. 19, pp. 780–785, July 1997.
[11] ONYX Digital EEG User’s Manual. Heemstede, The Netherlands:
patterns cannot be accurately estimated through traditional SBI BV, 1998.
visual inspection of video recordings. We intend to search [12] G. Johannsen and J. Bille, “A threshold selection method using informa-
for more accurate quantification methods, namely to employ tion measures,” in Proc. 6th Int. Conf. Pattern Recognition, 1982, vol.
1, pp. 140–142.
multicamera video acquisition to extract three-dimensional [13] M. Shah, K. Rangarajan, and P. S. Tsai, “Motion trajectories,” IEEE
information from which we may obtain deepness and also Trans. Syst., Man, Cybern., vol. 23, pp. 1138–1150, July-Aug. 1993.
reduce the cases of occlusion even in cases where a patient [14] A. M. Tekalp, Digital Video Processing. Englewood Cliffs, NJ: Pren-
tice Hall, 1995.
lies sideways or in a prone position. We were able, apparently [15] A. Jobbágy, “Fault diagnosis in biomedical engineering application,” ,
for the first time, to observe the course of the motions of Dept. Meas. Inform. Syst., Tech. Univ. Budapest, Budapest, Hungary,
Tech. Rep., 1997.
an epileptic seizure in a quantified way. This constitutes an [16] K. Rangarjan and M. Shah, “Establishing motion correspondence,”
important result of our work. We are now able to register CVGIP: Image Understanding, vol. 54, no. 1, pp. 56–73, 1991.
movements of different parts of the convulsive human body [17] I. K. Sethi and R. Jain, “Finding trajectories of feature points in a monoc-
ular image sequence,” IEEE Trans. Pattern Anal. Machine Intell., vol.
and quantify their speed, quantity of movement and 2-D PAMI-9, pp. 56–73, Jan. 1987.
spatial motion patterns. This type of quantification may make [18] R. F. Rashid, “Toward a system for the interpretation of moving light
a lot of difference in clinical epileptology. We hope this new displays,” IEEE Trans. Pattern Anal. Machine Intell., vol. PAMI-2, pp.
574–581, June 1980.
approach will aid clinicians to find more valuable evidences
that may lead to a more reliable seizure classification based on
quantified motor semiology. One immediate application can
be the help in dissociating epileptic from simulated seizures Zhanjian Li received the B.S. degree in electronic
(also called pseudoseizures). The possibility to correlate this engineering from Shanghai Jiaotong University,
Shanghai, China, in 1987; the M.S. degree in
quantified motion information with the EEG events occurring computer science from Shenyang Institute of Com-
in the course of the seizure or during small movements opens puting Technology (SICT), the Chinese Academy
a completely new research item that we intend to explore in of Sciences (CAS), Shenyang, China, in 1993;
the Ph.D. degree in electronics engineering and
the near future. Finally, this technique has a lot of potential Telecommunications from the University of Aveiro,
to study other human pathological states, namely movement Portugal, in 2002.
disorders or behavioral changes which are accompanied by He is currently a research staff at the Institute of
motor components. Electronics Engineering and Telematics of Aveiro
(IEETA), Aveiro, Portugal. His research interests include image analysis,
For an interactive overview of our work, readers can access object-oriented programming, and multimedia applications.
our web site at http://www.ieeta.pt/sias. Dr. Li received the president award from CAS in 1993.

António Martins da Silva received the medical de- Joao Paulo Silva Cunha (S’87–M’90) was born
gree and the Ph.D. degree on neurophysiology, from in Porto, Portugal. He received an Electronics
University of Porto, Portugal. and Telecommunications Engineering Diploma in
He is Professor of Human Physiology with the 1989 and the Ph.D. degree in 1996, both from the
Medical School at the Biomedical Sciences Institute University of Aveiro, Portugal.
“Abel Salazar,” University of Porto. He is the In 1996, he joined the Department of Electronics
Director of Clinical Neurophysiology Department, and Telecommunications at the University of Aviero,
Santo António Hospital, Porto, and Director of the where he is currently an Associate Professor. Previ-
Research Unit on Neuro and Psychophysiology ously, he was a Research Assistant at the Institute
IBMC, University of Porto. His current research of Biomedical Sciences, University of Porto, Porto,
fields/clinical work are mainly on epilepsy, sleep, Portugal. He is an invited Professor in the Masters in
and brain function. Biomedical Engineering program of the same University. He has been a Vis-
iting Scientist at the University of Florida, Gainsville, FL. and at the “Instituut
voor Epilepsiebestrijding,” Hemestede, The Netherlands. His research has been
devoted to the study of computer-based systems to support diagnosis in clinical
neurophysiology. His other current research interests are telemedicine and mul-
timedia healthcare information systems.
Dr. Cunha is member at the CEN/TC251 normalization body, a member of
the IEEE EMBS, vice-president of the Portuguese Association for Medical In-
formatics, and a member of the Portuguese Biomedical Engineering Society and
of the Portuguese League Against Epilepsy.